mm/backing-dev.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2
   3 #include <linux/wait.h>
   4 #include <linux/rbtree.h>
   5 #include <linux/backing-dev.h>
   6 #include <linux/kthread.h>
   7 #include <linux/freezer.h>
   8 #include <linux/fs.h>
   9 #include <linux/pagemap.h>
  10 #include <linux/mm.h>
  11 #include <linux/sched.h>
  12 #include <linux/module.h>
  13 #include <linux/writeback.h>
  14 #include <linux/device.h>
  15 #include <trace/events/writeback.h>
  16
  17 struct backing_dev_info noop_backing_dev_info = {
  18         .name           = "noop",
  19         .capabilities   = BDI_CAP_NO_ACCT_AND_WRITEBACK,
  20 };
  21 EXPORT_SYMBOL_GPL(noop_backing_dev_info);
  22
  23 static struct class *bdi_class;
  24 const char *bdi_unknown_name = "(unknown)";
  25
  26 /*
  27  * bdi_lock protects bdi_tree and updates to bdi_list. bdi_list has RCU
  28  * reader side locking.
  29  */
  30 DEFINE_SPINLOCK(bdi_lock);
  31 static u64 bdi_id_cursor;
  32 static struct rb_root bdi_tree = RB_ROOT;
  33 LIST_HEAD(bdi_list);
  34
  35 /* bdi_wq serves all asynchronous writeback tasks */
  36 struct workqueue_struct *bdi_wq;
  37
  38 #ifdef CONFIG_DEBUG_FS
  39 #include <linux/debugfs.h>
  40 #include <linux/seq_file.h>
  41
  42 static struct dentry *bdi_debug_root;
  43
  44 static void bdi_debug_init(void)
  45 {
  46         bdi_debug_root = debugfs_create_dir("bdi", NULL);
  47 }
  48
  49 static int bdi_debug_stats_show(struct seq_file *m, void *v)
  50 {
  51         struct backing_dev_info *bdi = m->private;
  52         struct bdi_writeback *wb = &bdi->wb;
  53         unsigned long background_thresh;
  54         unsigned long dirty_thresh;
  55         unsigned long wb_thresh;
  56         unsigned long nr_dirty, nr_io, nr_more_io, nr_dirty_time;
  57         struct inode *inode;
  58
  59         nr_dirty = nr_io = nr_more_io = nr_dirty_time = 0;
  60         spin_lock(&wb->list_lock);
  61         list_for_each_entry(inode, &wb->b_dirty, i_io_list)
  62                 nr_dirty++;
  63         list_for_each_entry(inode, &wb->b_io, i_io_list)
  64                 nr_io++;
  65         list_for_each_entry(inode, &wb->b_more_io, i_io_list)
  66                 nr_more_io++;
  67         list_for_each_entry(inode, &wb->b_dirty_time, i_io_list)
  68                 if (inode->i_state & I_DIRTY_TIME)
  69                         nr_dirty_time++;
  70         spin_unlock(&wb->list_lock);
  71
  72         global_dirty_limits(&background_thresh, &dirty_thresh);
  73         wb_thresh = wb_calc_thresh(wb, dirty_thresh);
  74
  75 #define K(x) ((x) << (PAGE_SHIFT - 10))
  76         seq_printf(m,
  77                    "BdiWriteback:       %10lu kB\n"
  78                    "BdiReclaimable:     %10lu kB\n"
  79                    "BdiDirtyThresh:     %10lu kB\n"
  80                    "DirtyThresh:        %10lu kB\n"
  81                    "BackgroundThresh:   %10lu kB\n"
  82                    "BdiDirtied:         %10lu kB\n"
  83                    "BdiWritten:         %10lu kB\n"
  84                    "BdiWriteBandwidth:  %10lu kBps\n"
  85                    "b_dirty:            %10lu\n"
  86                    "b_io:               %10lu\n"
  87                    "b_more_io:          %10lu\n"
  88                    "b_dirty_time:       %10lu\n"
  89                    "bdi_list:           %10u\n"
  90                    "state:              %10lx\n",
  91                    (unsigned long) K(wb_stat(wb, WB_WRITEBACK)),
  92                    (unsigned long) K(wb_stat(wb, WB_RECLAIMABLE)),
  93                    K(wb_thresh),
  94                    K(dirty_thresh),
  95                    K(background_thresh),
  96                    (unsigned long) K(wb_stat(wb, WB_DIRTIED)),
  97                    (unsigned long) K(wb_stat(wb, WB_WRITTEN)),
  98                    (unsigned long) K(wb->write_bandwidth),
  99                    nr_dirty,
 100                    nr_io,
 101                    nr_more_io,
 102                    nr_dirty_time,
 103                    !list_empty(&bdi->bdi_list), bdi->wb.state);
 104 #undef K
 105
 106         return 0;
 107 }
 108 DEFINE_SHOW_ATTRIBUTE(bdi_debug_stats);
 109
 110 static void bdi_debug_register(struct backing_dev_info *bdi, const char *name)
 111 {
 112         bdi->debug_dir = debugfs_create_dir(name, bdi_debug_root);
 113
 114         debugfs_create_file("stats", 0444, bdi->debug_dir, bdi,
 115                             &bdi_debug_stats_fops);
 116 }
 117
 118 static void bdi_debug_unregister(struct backing_dev_info *bdi)
 119 {
 120         debugfs_remove_recursive(bdi->debug_dir);
 121 }
 122 #else
 123 static inline void bdi_debug_init(void)
 124 {
 125 }
 126 static inline void bdi_debug_register(struct backing_dev_info *bdi,
 127                                       const char *name)
 128 {
 129 }
 130 static inline void bdi_debug_unregister(struct backing_dev_info *bdi)
 131 {
 132 }
 133 #endif
 134
 135 static ssize_t read_ahead_kb_store(struct device *dev,
 136                                   struct device_attribute *attr,
 137                                   const char *buf, size_t count)
 138 {
 139         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 140         unsigned long read_ahead_kb;
 141         ssize_t ret;
 142
 143         ret = kstrtoul(buf, 10, &read_ahead_kb);
 144         if (ret < 0)
 145                 return ret;
 146
 147         bdi->ra_pages = read_ahead_kb >> (PAGE_SHIFT - 10);
 148
 149         return count;
 150 }
 151
 152 #define K(pages) ((pages) << (PAGE_SHIFT - 10))
 153
 154 #define BDI_SHOW(name, expr)                                            \
 155 static ssize_t name##_show(struct device *dev,                          \
 156                            struct device_attribute *attr, char *page)   \
 157 {                                                                       \
 158         struct backing_dev_info *bdi = dev_get_drvdata(dev);            \
 159                                                                         \
 160         return snprintf(page, PAGE_SIZE-1, "%lld\n", (long long)expr);  \
 161 }                                                                       \
 162 static DEVICE_ATTR_RW(name);
 163
 164 BDI_SHOW(read_ahead_kb, K(bdi->ra_pages))
 165
 166 static ssize_t min_ratio_store(struct device *dev,
 167                 struct device_attribute *attr, const char *buf, size_t count)
 168 {
 169         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 170         unsigned int ratio;
 171         ssize_t ret;
 172
 173         ret = kstrtouint(buf, 10, &ratio);
 174         if (ret < 0)
 175                 return ret;
 176
 177         ret = bdi_set_min_ratio(bdi, ratio);
 178         if (!ret)
 179                 ret = count;
 180
 181         return ret;
 182 }
 183 BDI_SHOW(min_ratio, bdi->min_ratio)
 184
 185 static ssize_t max_ratio_store(struct device *dev,
 186                 struct device_attribute *attr, const char *buf, size_t count)
 187 {
 188         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 189         unsigned int ratio;
 190         ssize_t ret;
 191
 192         ret = kstrtouint(buf, 10, &ratio);
 193         if (ret < 0)
 194                 return ret;
 195
 196         ret = bdi_set_max_ratio(bdi, ratio);
 197         if (!ret)
 198                 ret = count;
 199
 200         return ret;
 201 }
 202 BDI_SHOW(max_ratio, bdi->max_ratio)
 203
 204 static ssize_t stable_pages_required_show(struct device *dev,
 205                                           struct device_attribute *attr,
 206                                           char *page)
 207 {
 208         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 209
 210         return snprintf(page, PAGE_SIZE-1, "%d\n",
 211                         bdi_cap_stable_pages_required(bdi) ? 1 : 0);
 212 }
 213 static DEVICE_ATTR_RO(stable_pages_required);
 214
 215 static struct attribute *bdi_dev_attrs[] = {
 216         &dev_attr_read_ahead_kb.attr,
 217         &dev_attr_min_ratio.attr,
 218         &dev_attr_max_ratio.attr,
 219         &dev_attr_stable_pages_required.attr,
 220         NULL,
 221 };
 222 ATTRIBUTE_GROUPS(bdi_dev);
 223
 224 static __init int bdi_class_init(void)
 225 {
 226         bdi_class = class_create(THIS_MODULE, "bdi");
 227         if (IS_ERR(bdi_class))
 228                 return PTR_ERR(bdi_class);
 229
 230         bdi_class->dev_groups = bdi_dev_groups;
 231         bdi_debug_init();
 232
 233         return 0;
 234 }
 235 postcore_initcall(bdi_class_init);
 236
 237 static int bdi_init(struct backing_dev_info *bdi);
 238
 239 static int __init default_bdi_init(void)
 240 {
 241         int err;
 242
 243         bdi_wq = alloc_workqueue("writeback", WQ_MEM_RECLAIM | WQ_UNBOUND |
 244                                  WQ_SYSFS, 0);
 245         if (!bdi_wq)
 246                 return -ENOMEM;
 247
 248         err = bdi_init(&noop_backing_dev_info);
 249
 250         return err;
 251 }
 252 subsys_initcall(default_bdi_init);
 253
 254 /*
 255  * This function is used when the first inode for this wb is marked dirty. It
 256  * wakes-up the corresponding bdi thread which should then take care of the
 257  * periodic background write-out of dirty inodes. Since the write-out would
 258  * starts only 'dirty_writeback_interval' centisecs from now anyway, we just
 259  * set up a timer which wakes the bdi thread up later.
 260  *
 261  * Note, we wouldn't bother setting up the timer, but this function is on the
 262  * fast-path (used by '__mark_inode_dirty()'), so we save few context switches
 263  * by delaying the wake-up.
 264  *
 265  * We have to be careful not to postpone flush work if it is scheduled for
 266  * earlier. Thus we use queue_delayed_work().
 267  */
 268 void wb_wakeup_delayed(struct bdi_writeback *wb)
 269 {
 270         unsigned long timeout;
 271
 272         timeout = msecs_to_jiffies(dirty_writeback_interval * 10);
 273         spin_lock_bh(&wb->work_lock);
 274         if (test_bit(WB_registered, &wb->state))
 275                 queue_delayed_work(bdi_wq, &wb->dwork, timeout);
 276         spin_unlock_bh(&wb->work_lock);
 277 }
 278
 279 /*
 280  * Initial write bandwidth: 100 MB/s
 281  */
 282 #define INIT_BW         (100 << (20 - PAGE_SHIFT))
 283
 284 static int wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi,
 285                    int blkcg_id, gfp_t gfp)
 286 {
 287         int i, err;
 288
 289         memset(wb, 0, sizeof(*wb));
 290
 291         if (wb != &bdi->wb)
 292                 bdi_get(bdi);
 293         wb->bdi = bdi;
 294         wb->last_old_flush = jiffies;
 295         INIT_LIST_HEAD(&wb->b_dirty);
 296         INIT_LIST_HEAD(&wb->b_io);
 297         INIT_LIST_HEAD(&wb->b_more_io);
 298         INIT_LIST_HEAD(&wb->b_dirty_time);
 299         spin_lock_init(&wb->list_lock);
 300
 301         wb->bw_time_stamp = jiffies;
 302         wb->balanced_dirty_ratelimit = INIT_BW;
 303         wb->dirty_ratelimit = INIT_BW;
 304         wb->write_bandwidth = INIT_BW;
 305         wb->avg_write_bandwidth = INIT_BW;
 306
 307         spin_lock_init(&wb->work_lock);
 308         INIT_LIST_HEAD(&wb->work_list);
 309         INIT_DELAYED_WORK(&wb->dwork, wb_workfn);
 310         wb->dirty_sleep = jiffies;
 311
 312         wb->congested = wb_congested_get_create(bdi, blkcg_id, gfp);
 313         if (!wb->congested) {
 314                 err = -ENOMEM;
 315                 goto out_put_bdi;
 316         }
 317
 318         err = fprop_local_init_percpu(&wb->completions, gfp);
 319         if (err)
 320                 goto out_put_cong;
 321
 322         for (i = 0; i < NR_WB_STAT_ITEMS; i++) {
 323                 err = percpu_counter_init(&wb->stat[i], 0, gfp);
 324                 if (err)
 325                         goto out_destroy_stat;
 326         }
 327
 328         return 0;
 329
 330 out_destroy_stat:
 331         while (i--)
 332                 percpu_counter_destroy(&wb->stat[i]);
 333         fprop_local_destroy_percpu(&wb->completions);
 334 out_put_cong:
 335         wb_congested_put(wb->congested);
 336 out_put_bdi:
 337         if (wb != &bdi->wb)
 338                 bdi_put(bdi);
 339         return err;
 340 }
 341
 342 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb);
 343
 344 /*
 345  * Remove bdi from the global list and shutdown any threads we have running
 346  */
 347 static void wb_shutdown(struct bdi_writeback *wb)
 348 {
 349         /* Make sure nobody queues further work */
 350         spin_lock_bh(&wb->work_lock);
 351         if (!test_and_clear_bit(WB_registered, &wb->state)) {
 352                 spin_unlock_bh(&wb->work_lock);
 353                 return;
 354         }
 355         spin_unlock_bh(&wb->work_lock);
 356
 357         cgwb_remove_from_bdi_list(wb);
 358         /*
 359          * Drain work list and shutdown the delayed_work.  !WB_registered
 360          * tells wb_workfn() that @wb is dying and its work_list needs to
 361          * be drained no matter what.
 362          */
 363         mod_delayed_work(bdi_wq, &wb->dwork, 0);
 364         flush_delayed_work(&wb->dwork);
 365         WARN_ON(!list_empty(&wb->work_list));
 366 }
 367
 368 static void wb_exit(struct bdi_writeback *wb)
 369 {
 370         int i;
 371
 372         WARN_ON(delayed_work_pending(&wb->dwork));
 373
 374         for (i = 0; i < NR_WB_STAT_ITEMS; i++)
 375                 percpu_counter_destroy(&wb->stat[i]);
 376
 377         fprop_local_destroy_percpu(&wb->completions);
 378         wb_congested_put(wb->congested);
 379         if (wb != &wb->bdi->wb)
 380                 bdi_put(wb->bdi);
 381 }
 382
 383 #ifdef CONFIG_CGROUP_WRITEBACK
 384
 385 #include <linux/memcontrol.h>
 386
 387 /*
 388  * cgwb_lock protects bdi->cgwb_tree, bdi->cgwb_congested_tree,
 389  * blkcg->cgwb_list, and memcg->cgwb_list.  bdi->cgwb_tree is also RCU
 390  * protected.
 391  */
 392 static DEFINE_SPINLOCK(cgwb_lock);
 393 static struct workqueue_struct *cgwb_release_wq;
 394
 395 /**
 396  * wb_congested_get_create - get or create a wb_congested
 397  * @bdi: associated bdi
 398  * @blkcg_id: ID of the associated blkcg
 399  * @gfp: allocation mask
 400  *
 401  * Look up the wb_congested for @blkcg_id on @bdi.  If missing, create one.
 402  * The returned wb_congested has its reference count incremented.  Returns
 403  * NULL on failure.
 404  */
 405 struct bdi_writeback_congested *
 406 wb_congested_get_create(struct backing_dev_info *bdi, int blkcg_id, gfp_t gfp)
 407 {
 408         struct bdi_writeback_congested *new_congested = NULL, *congested;
 409         struct rb_node **node, *parent;
 410         unsigned long flags;
 411 retry:
 412         spin_lock_irqsave(&cgwb_lock, flags);
 413
 414         node = &bdi->cgwb_congested_tree.rb_node;
 415         parent = NULL;
 416
 417         while (*node != NULL) {
 418                 parent = *node;
 419                 congested = rb_entry(parent, struct bdi_writeback_congested,
 420                                      rb_node);
 421                 if (congested->blkcg_id < blkcg_id)
 422                         node = &parent->rb_left;
 423                 else if (congested->blkcg_id > blkcg_id)
 424                         node = &parent->rb_right;
 425                 else
 426                         goto found;
 427         }
 428
 429         if (new_congested) {
 430                 /* !found and storage for new one already allocated, insert */
 431                 congested = new_congested;
 432                 rb_link_node(&congested->rb_node, parent, node);
 433                 rb_insert_color(&congested->rb_node, &bdi->cgwb_congested_tree);
 434                 spin_unlock_irqrestore(&cgwb_lock, flags);
 435                 return congested;
 436         }
 437
 438         spin_unlock_irqrestore(&cgwb_lock, flags);
 439
 440         /* allocate storage for new one and retry */
 441         new_congested = kzalloc(sizeof(*new_congested), gfp);
 442         if (!new_congested)
 443                 return NULL;
 444
 445         refcount_set(&new_congested->refcnt, 1);
 446         new_congested->__bdi = bdi;
 447         new_congested->blkcg_id = blkcg_id;
 448         goto retry;
 449
 450 found:
 451         refcount_inc(&congested->refcnt);
 452         spin_unlock_irqrestore(&cgwb_lock, flags);
 453         kfree(new_congested);
 454         return congested;
 455 }
 456
 457 /**
 458  * wb_congested_put - put a wb_congested
 459  * @congested: wb_congested to put
 460  *
 461  * Put @congested and destroy it if the refcnt reaches zero.
 462  */
 463 void wb_congested_put(struct bdi_writeback_congested *congested)
 464 {
 465         unsigned long flags;
 466
 467         if (!refcount_dec_and_lock_irqsave(&congested->refcnt, &cgwb_lock, &flags))
 468                 return;
 469
 470         /* bdi might already have been destroyed leaving @congested unlinked */
 471         if (congested->__bdi) {
 472                 rb_erase(&congested->rb_node,
 473                          &congested->__bdi->cgwb_congested_tree);
 474                 congested->__bdi = NULL;
 475         }
 476
 477         spin_unlock_irqrestore(&cgwb_lock, flags);
 478         kfree(congested);
 479 }
 480
 481 static void cgwb_release_workfn(struct work_struct *work)
 482 {
 483         struct bdi_writeback *wb = container_of(work, struct bdi_writeback,
 484                                                 release_work);
 485         struct blkcg *blkcg = css_to_blkcg(wb->blkcg_css);
 486
 487         mutex_lock(&wb->bdi->cgwb_release_mutex);
 488         wb_shutdown(wb);
 489
 490         css_put(wb->memcg_css);
 491         css_put(wb->blkcg_css);
 492         mutex_unlock(&wb->bdi->cgwb_release_mutex);
 493
 494         /* triggers blkg destruction if cgwb_refcnt becomes zero */
 495         blkcg_cgwb_put(blkcg);
 496
 497         fprop_local_destroy_percpu(&wb->memcg_completions);
 498         percpu_ref_exit(&wb->refcnt);
 499         wb_exit(wb);
 500         kfree_rcu(wb, rcu);
 501 }
 502
 503 static void cgwb_release(struct percpu_ref *refcnt)
 504 {
 505         struct bdi_writeback *wb = container_of(refcnt, struct bdi_writeback,
 506                                                 refcnt);
 507         queue_work(cgwb_release_wq, &wb->release_work);
 508 }
 509
 510 static void cgwb_kill(struct bdi_writeback *wb)
 511 {
 512         lockdep_assert_held(&cgwb_lock);
 513
 514         WARN_ON(!radix_tree_delete(&wb->bdi->cgwb_tree, wb->memcg_css->id));
 515         list_del(&wb->memcg_node);
 516         list_del(&wb->blkcg_node);
 517         percpu_ref_kill(&wb->refcnt);
 518 }
 519
 520 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
 521 {
 522         spin_lock_irq(&cgwb_lock);
 523         list_del_rcu(&wb->bdi_node);
 524         spin_unlock_irq(&cgwb_lock);
 525 }
 526
 527 static int cgwb_create(struct backing_dev_info *bdi,
 528                        struct cgroup_subsys_state *memcg_css, gfp_t gfp)
 529 {
 530         struct mem_cgroup *memcg;
 531         struct cgroup_subsys_state *blkcg_css;
 532         struct blkcg *blkcg;
 533         struct list_head *memcg_cgwb_list, *blkcg_cgwb_list;
 534         struct bdi_writeback *wb;
 535         unsigned long flags;
 536         int ret = 0;
 537
 538         memcg = mem_cgroup_from_css(memcg_css);
 539         blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
 540         blkcg = css_to_blkcg(blkcg_css);
 541         memcg_cgwb_list = &memcg->cgwb_list;
 542         blkcg_cgwb_list = &blkcg->cgwb_list;
 543
 544         /* look up again under lock and discard on blkcg mismatch */
 545         spin_lock_irqsave(&cgwb_lock, flags);
 546         wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
 547         if (wb && wb->blkcg_css != blkcg_css) {
 548                 cgwb_kill(wb);
 549                 wb = NULL;
 550         }
 551         spin_unlock_irqrestore(&cgwb_lock, flags);
 552         if (wb)
 553                 goto out_put;
 554
 555         /* need to create a new one */
 556         wb = kmalloc(sizeof(*wb), gfp);
 557         if (!wb) {
 558                 ret = -ENOMEM;
 559                 goto out_put;
 560         }
 561
 562         ret = wb_init(wb, bdi, blkcg_css->id, gfp);
 563         if (ret)
 564                 goto err_free;
 565
 566         ret = percpu_ref_init(&wb->refcnt, cgwb_release, 0, gfp);
 567         if (ret)
 568                 goto err_wb_exit;
 569
 570         ret = fprop_local_init_percpu(&wb->memcg_completions, gfp);
 571         if (ret)
 572                 goto err_ref_exit;
 573
 574         wb->memcg_css = memcg_css;
 575         wb->blkcg_css = blkcg_css;
 576         INIT_WORK(&wb->release_work, cgwb_release_workfn);
 577         set_bit(WB_registered, &wb->state);
 578
 579         /*
 580          * The root wb determines the registered state of the whole bdi and
 581          * memcg_cgwb_list and blkcg_cgwb_list's next pointers indicate
 582          * whether they're still online.  Don't link @wb if any is dead.
 583          * See wb_memcg_offline() and wb_blkcg_offline().
 584          */
 585         ret = -ENODEV;
 586         spin_lock_irqsave(&cgwb_lock, flags);
 587         if (test_bit(WB_registered, &bdi->wb.state) &&
 588             blkcg_cgwb_list->next && memcg_cgwb_list->next) {
 589                 /* we might have raced another instance of this function */
 590                 ret = radix_tree_insert(&bdi->cgwb_tree, memcg_css->id, wb);
 591                 if (!ret) {
 592                         list_add_tail_rcu(&wb->bdi_node, &bdi->wb_list);
 593                         list_add(&wb->memcg_node, memcg_cgwb_list);
 594                         list_add(&wb->blkcg_node, blkcg_cgwb_list);
 595                         blkcg_cgwb_get(blkcg);
 596                         css_get(memcg_css);
 597                         css_get(blkcg_css);
 598                 }
 599         }
 600         spin_unlock_irqrestore(&cgwb_lock, flags);
 601         if (ret) {
 602                 if (ret == -EEXIST)
 603                         ret = 0;
 604                 goto err_fprop_exit;
 605         }
 606         goto out_put;
 607
 608 err_fprop_exit:
 609         fprop_local_destroy_percpu(&wb->memcg_completions);
 610 err_ref_exit:
 611         percpu_ref_exit(&wb->refcnt);
 612 err_wb_exit:
 613         wb_exit(wb);
 614 err_free:
 615         kfree(wb);
 616 out_put:
 617         css_put(blkcg_css);
 618         return ret;
 619 }
 620
 621 /**
 622  * wb_get_lookup - get wb for a given memcg
 623  * @bdi: target bdi
 624  * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
 625  *
 626  * Try to get the wb for @memcg_css on @bdi.  The returned wb has its
 627  * refcount incremented.
 628  *
 629  * This function uses css_get() on @memcg_css and thus expects its refcnt
 630  * to be positive on invocation.  IOW, rcu_read_lock() protection on
 631  * @memcg_css isn't enough.  try_get it before calling this function.
 632  *
 633  * A wb is keyed by its associated memcg.  As blkcg implicitly enables
 634  * memcg on the default hierarchy, memcg association is guaranteed to be
 635  * more specific (equal or descendant to the associated blkcg) and thus can
 636  * identify both the memcg and blkcg associations.
 637  *
 638  * Because the blkcg associated with a memcg may change as blkcg is enabled
 639  * and disabled closer to root in the hierarchy, each wb keeps track of
 640  * both the memcg and blkcg associated with it and verifies the blkcg on
 641  * each lookup.  On mismatch, the existing wb is discarded and a new one is
 642  * created.
 643  */
 644 struct bdi_writeback *wb_get_lookup(struct backing_dev_info *bdi,
 645                                     struct cgroup_subsys_state *memcg_css)
 646 {
 647         struct bdi_writeback *wb;
 648
 649         if (!memcg_css->parent)
 650                 return &bdi->wb;
 651
 652         rcu_read_lock();
 653         wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
 654         if (wb) {
 655                 struct cgroup_subsys_state *blkcg_css;
 656
 657                 /* see whether the blkcg association has changed */
 658                 blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
 659                 if (unlikely(wb->blkcg_css != blkcg_css || !wb_tryget(wb)))
 660                         wb = NULL;
 661                 css_put(blkcg_css);
 662         }
 663         rcu_read_unlock();
 664
 665         return wb;
 666 }
 667
 668 /**
 669  * wb_get_create - get wb for a given memcg, create if necessary
 670  * @bdi: target bdi
 671  * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
 672  * @gfp: allocation mask to use
 673  *
 674  * Try to get the wb for @memcg_css on @bdi.  If it doesn't exist, try to
 675  * create one.  See wb_get_lookup() for more details.
 676  */
 677 struct bdi_writeback *wb_get_create(struct backing_dev_info *bdi,
 678                                     struct cgroup_subsys_state *memcg_css,
 679                                     gfp_t gfp)
 680 {
 681         struct bdi_writeback *wb;
 682
 683         might_sleep_if(gfpflags_allow_blocking(gfp));
 684
 685         if (!memcg_css->parent)
 686                 return &bdi->wb;
 687
 688         do {
 689                 wb = wb_get_lookup(bdi, memcg_css);
 690         } while (!wb && !cgwb_create(bdi, memcg_css, gfp));
 691
 692         return wb;
 693 }
 694
 695 static int cgwb_bdi_init(struct backing_dev_info *bdi)
 696 {
 697         int ret;
 698
 699         INIT_RADIX_TREE(&bdi->cgwb_tree, GFP_ATOMIC);
 700         bdi->cgwb_congested_tree = RB_ROOT;
 701         mutex_init(&bdi->cgwb_release_mutex);
 702         init_rwsem(&bdi->wb_switch_rwsem);
 703
 704         ret = wb_init(&bdi->wb, bdi, 1, GFP_KERNEL);
 705         if (!ret) {
 706                 bdi->wb.memcg_css = &root_mem_cgroup->css;
 707                 bdi->wb.blkcg_css = blkcg_root_css;
 708         }
 709         return ret;
 710 }
 711
 712 static void cgwb_bdi_unregister(struct backing_dev_info *bdi)
 713 {
 714         struct radix_tree_iter iter;
 715         void **slot;
 716         struct bdi_writeback *wb;
 717
 718         WARN_ON(test_bit(WB_registered, &bdi->wb.state));
 719
 720         spin_lock_irq(&cgwb_lock);
 721         radix_tree_for_each_slot(slot, &bdi->cgwb_tree, &iter, 0)
 722                 cgwb_kill(*slot);
 723         spin_unlock_irq(&cgwb_lock);
 724
 725         mutex_lock(&bdi->cgwb_release_mutex);
 726         spin_lock_irq(&cgwb_lock);
 727         while (!list_empty(&bdi->wb_list)) {
 728                 wb = list_first_entry(&bdi->wb_list, struct bdi_writeback,
 729                                       bdi_node);
 730                 spin_unlock_irq(&cgwb_lock);
 731                 wb_shutdown(wb);
 732                 spin_lock_irq(&cgwb_lock);
 733         }
 734         spin_unlock_irq(&cgwb_lock);
 735         mutex_unlock(&bdi->cgwb_release_mutex);
 736 }
 737
 738 /**
 739  * wb_memcg_offline - kill all wb's associated with a memcg being offlined
 740  * @memcg: memcg being offlined
 741  *
 742  * Also prevents creation of any new wb's associated with @memcg.
 743  */
 744 void wb_memcg_offline(struct mem_cgroup *memcg)
 745 {
 746         struct list_head *memcg_cgwb_list = &memcg->cgwb_list;
 747         struct bdi_writeback *wb, *next;
 748
 749         spin_lock_irq(&cgwb_lock);
 750         list_for_each_entry_safe(wb, next, memcg_cgwb_list, memcg_node)
 751                 cgwb_kill(wb);
 752         memcg_cgwb_list->next = NULL;   /* prevent new wb's */
 753         spin_unlock_irq(&cgwb_lock);
 754 }
 755
 756 /**
 757  * wb_blkcg_offline - kill all wb's associated with a blkcg being offlined
 758  * @blkcg: blkcg being offlined
 759  *
 760  * Also prevents creation of any new wb's associated with @blkcg.
 761  */
 762 void wb_blkcg_offline(struct blkcg *blkcg)
 763 {
 764         struct bdi_writeback *wb, *next;
 765
 766         spin_lock_irq(&cgwb_lock);
 767         list_for_each_entry_safe(wb, next, &blkcg->cgwb_list, blkcg_node)
 768                 cgwb_kill(wb);
 769         blkcg->cgwb_list.next = NULL;   /* prevent new wb's */
 770         spin_unlock_irq(&cgwb_lock);
 771 }
 772
 773 static void cgwb_bdi_exit(struct backing_dev_info *bdi)
 774 {
 775         struct rb_node *rbn;
 776
 777         spin_lock_irq(&cgwb_lock);
 778         while ((rbn = rb_first(&bdi->cgwb_congested_tree))) {
 779                 struct bdi_writeback_congested *congested =
 780                         rb_entry(rbn, struct bdi_writeback_congested, rb_node);
 781
 782                 rb_erase(rbn, &bdi->cgwb_congested_tree);
 783                 congested->__bdi = NULL;        /* mark @congested unlinked */
 784         }
 785         spin_unlock_irq(&cgwb_lock);
 786 }
 787
 788 static void cgwb_bdi_register(struct backing_dev_info *bdi)
 789 {
 790         spin_lock_irq(&cgwb_lock);
 791         list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
 792         spin_unlock_irq(&cgwb_lock);
 793 }
 794
 795 static int __init cgwb_init(void)
 796 {
 797         /*
 798          * There can be many concurrent release work items overwhelming
 799          * system_wq.  Put them in a separate wq and limit concurrency.
 800          * There's no point in executing many of these in parallel.
 801          */
 802         cgwb_release_wq = alloc_workqueue("cgwb_release", 0, 1);
 803         if (!cgwb_release_wq)
 804                 return -ENOMEM;
 805
 806         return 0;
 807 }
 808 subsys_initcall(cgwb_init);
 809
 810 #else   /* CONFIG_CGROUP_WRITEBACK */
 811
 812 static int cgwb_bdi_init(struct backing_dev_info *bdi)
 813 {
 814         int err;
 815
 816         bdi->wb_congested = kzalloc(sizeof(*bdi->wb_congested), GFP_KERNEL);
 817         if (!bdi->wb_congested)
 818                 return -ENOMEM;
 819
 820         refcount_set(&bdi->wb_congested->refcnt, 1);
 821
 822         err = wb_init(&bdi->wb, bdi, 1, GFP_KERNEL);
 823         if (err) {
 824                 wb_congested_put(bdi->wb_congested);
 825                 return err;
 826         }
 827         return 0;
 828 }
 829
 830 static void cgwb_bdi_unregister(struct backing_dev_info *bdi) { }
 831
 832 static void cgwb_bdi_exit(struct backing_dev_info *bdi)
 833 {
 834         wb_congested_put(bdi->wb_congested);
 835 }
 836
 837 static void cgwb_bdi_register(struct backing_dev_info *bdi)
 838 {
 839         list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
 840 }
 841
 842 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
 843 {
 844         list_del_rcu(&wb->bdi_node);
 845 }
 846
 847 #endif  /* CONFIG_CGROUP_WRITEBACK */
 848
 849 static int bdi_init(struct backing_dev_info *bdi)
 850 {
 851         int ret;
 852
 853         bdi->dev = NULL;
 854
 855         kref_init(&bdi->refcnt);
 856         bdi->min_ratio = 0;
 857         bdi->max_ratio = 100;
 858         bdi->max_prop_frac = FPROP_FRAC_BASE;
 859         INIT_LIST_HEAD(&bdi->bdi_list);
 860         INIT_LIST_HEAD(&bdi->wb_list);
 861         init_waitqueue_head(&bdi->wb_waitq);
 862
 863         ret = cgwb_bdi_init(bdi);
 864
 865         return ret;
 866 }
 867
 868 struct backing_dev_info *bdi_alloc_node(gfp_t gfp_mask, int node_id)
 869 {
 870         struct backing_dev_info *bdi;
 871
 872         bdi = kmalloc_node(sizeof(struct backing_dev_info),
 873                            gfp_mask | __GFP_ZERO, node_id);
 874         if (!bdi)
 875                 return NULL;
 876
 877         if (bdi_init(bdi)) {
 878                 kfree(bdi);
 879                 return NULL;
 880         }
 881         return bdi;
 882 }
 883 EXPORT_SYMBOL(bdi_alloc_node);
 884
 885 static struct rb_node **bdi_lookup_rb_node(u64 id, struct rb_node **parentp)
 886 {
 887         struct rb_node **p = &bdi_tree.rb_node;
 888         struct rb_node *parent = NULL;
 889         struct backing_dev_info *bdi;
 890
 891         lockdep_assert_held(&bdi_lock);
 892
 893         while (*p) {
 894                 parent = *p;
 895                 bdi = rb_entry(parent, struct backing_dev_info, rb_node);
 896
 897                 if (bdi->id > id)
 898                         p = &(*p)->rb_left;
 899                 else if (bdi->id < id)
 900                         p = &(*p)->rb_right;
 901                 else
 902                         break;
 903         }
 904
 905         if (parentp)
 906                 *parentp = parent;
 907         return p;
 908 }
 909
 910 /**
 911  * bdi_get_by_id - lookup and get bdi from its id
 912  * @id: bdi id to lookup
 913  *
 914  * Find bdi matching @id and get it.  Returns NULL if the matching bdi
 915  * doesn't exist or is already unregistered.
 916  */
 917 struct backing_dev_info *bdi_get_by_id(u64 id)
 918 {
 919         struct backing_dev_info *bdi = NULL;
 920         struct rb_node **p;
 921
 922         spin_lock_bh(&bdi_lock);
 923         p = bdi_lookup_rb_node(id, NULL);
 924         if (*p) {
 925                 bdi = rb_entry(*p, struct backing_dev_info, rb_node);
 926                 bdi_get(bdi);
 927         }
 928         spin_unlock_bh(&bdi_lock);
 929
 930         return bdi;
 931 }
 932
 933 int bdi_register_va(struct backing_dev_info *bdi, const char *fmt, va_list args)
 934 {
 935         struct device *dev;
 936         struct rb_node *parent, **p;
 937
 938         if (bdi->dev)   /* The driver needs to use separate queues per device */
 939                 return 0;
 940
 941         dev = device_create_vargs(bdi_class, NULL, MKDEV(0, 0), bdi, fmt, args);
 942         if (IS_ERR(dev))
 943                 return PTR_ERR(dev);
 944
 945         cgwb_bdi_register(bdi);
 946         bdi->dev = dev;
 947
 948         bdi_debug_register(bdi, dev_name(dev));
 949         set_bit(WB_registered, &bdi->wb.state);
 950
 951         spin_lock_bh(&bdi_lock);
 952
 953         bdi->id = ++bdi_id_cursor;
 954
 955         p = bdi_lookup_rb_node(bdi->id, &parent);
 956         rb_link_node(&bdi->rb_node, parent, p);
 957         rb_insert_color(&bdi->rb_node, &bdi_tree);
 958
 959         list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
 960
 961         spin_unlock_bh(&bdi_lock);
 962
 963         trace_writeback_bdi_register(bdi);
 964         return 0;
 965 }
 966 EXPORT_SYMBOL(bdi_register_va);
 967
 968 int bdi_register(struct backing_dev_info *bdi, const char *fmt, ...)
 969 {
 970         va_list args;
 971         int ret;
 972
 973         va_start(args, fmt);
 974         ret = bdi_register_va(bdi, fmt, args);
 975         va_end(args);
 976         return ret;
 977 }
 978 EXPORT_SYMBOL(bdi_register);
 979
 980 int bdi_register_owner(struct backing_dev_info *bdi, struct device *owner)
 981 {
 982         int rc;
 983
 984         rc = bdi_register(bdi, "%u:%u", MAJOR(owner->devt), MINOR(owner->devt));
 985         if (rc)
 986                 return rc;
 987         /* Leaking owner reference... */
 988         WARN_ON(bdi->owner);
 989         bdi->owner = owner;
 990         get_device(owner);
 991         return 0;
 992 }
 993 EXPORT_SYMBOL(bdi_register_owner);
 994
 995 /*
 996  * Remove bdi from bdi_list, and ensure that it is no longer visible
 997  */
 998 static void bdi_remove_from_list(struct backing_dev_info *bdi)
 999 {
1000         spin_lock_bh(&bdi_lock);
1001         rb_erase(&bdi->rb_node, &bdi_tree);
1002         list_del_rcu(&bdi->bdi_list);
1003         spin_unlock_bh(&bdi_lock);
1004
1005         synchronize_rcu_expedited();
1006 }
1007
1008 void bdi_unregister(struct backing_dev_info *bdi)
1009 {
1010         /* make sure nobody finds us on the bdi_list anymore */
1011         bdi_remove_from_list(bdi);
1012         wb_shutdown(&bdi->wb);
1013         cgwb_bdi_unregister(bdi);
1014
1015         if (bdi->dev) {
1016                 bdi_debug_unregister(bdi);
1017                 device_unregister(bdi->dev);
1018                 bdi->dev = NULL;
1019         }
1020
1021         if (bdi->owner) {
1022                 put_device(bdi->owner);
1023                 bdi->owner = NULL;
1024         }
1025 }
1026
1027 static void release_bdi(struct kref *ref)
1028 {
1029         struct backing_dev_info *bdi =
1030                         container_of(ref, struct backing_dev_info, refcnt);
1031
1032         if (test_bit(WB_registered, &bdi->wb.state))
1033                 bdi_unregister(bdi);
1034         WARN_ON_ONCE(bdi->dev);
1035         wb_exit(&bdi->wb);
1036         cgwb_bdi_exit(bdi);
1037         kfree(bdi);
1038 }
1039
1040 void bdi_put(struct backing_dev_info *bdi)
1041 {
1042         kref_put(&bdi->refcnt, release_bdi);
1043 }
1044 EXPORT_SYMBOL(bdi_put);
1045
1046 static wait_queue_head_t congestion_wqh[2] = {
1047                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]),
1048                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1])
1049         };
1050 static atomic_t nr_wb_congested[2];
1051
1052 void clear_wb_congested(struct bdi_writeback_congested *congested, int sync)
1053 {
1054         wait_queue_head_t *wqh = &congestion_wqh[sync];
1055         enum wb_congested_state bit;
1056
1057         bit = sync ? WB_sync_congested : WB_async_congested;
1058         if (test_and_clear_bit(bit, &congested->state))
1059                 atomic_dec(&nr_wb_congested[sync]);
1060         smp_mb__after_atomic();
1061         if (waitqueue_active(wqh))
1062                 wake_up(wqh);
1063 }
1064 EXPORT_SYMBOL(clear_wb_congested);
1065
1066 void set_wb_congested(struct bdi_writeback_congested *congested, int sync)
1067 {
1068         enum wb_congested_state bit;
1069
1070         bit = sync ? WB_sync_congested : WB_async_congested;
1071         if (!test_and_set_bit(bit, &congested->state))
1072                 atomic_inc(&nr_wb_congested[sync]);
1073 }
1074 EXPORT_SYMBOL(set_wb_congested);
1075
1076 /**
1077  * congestion_wait - wait for a backing_dev to become uncongested
1078  * @sync: SYNC or ASYNC IO
1079  * @timeout: timeout in jiffies
1080  *
1081  * Waits for up to @timeout jiffies for a backing_dev (any backing_dev) to exit
1082  * write congestion.  If no backing_devs are congested then just wait for the
1083  * next write to be completed.
1084  */
1085 long congestion_wait(int sync, long timeout)
1086 {
1087         long ret;
1088         unsigned long start = jiffies;
1089         DEFINE_WAIT(wait);
1090         wait_queue_head_t *wqh = &congestion_wqh[sync];
1091
1092         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1093         ret = io_schedule_timeout(timeout);
1094         finish_wait(wqh, &wait);
1095
1096         trace_writeback_congestion_wait(jiffies_to_usecs(timeout),
1097                                         jiffies_to_usecs(jiffies - start));
1098
1099         return ret;
1100 }
1101 EXPORT_SYMBOL(congestion_wait);
1102
1103 /**
1104  * wait_iff_congested - Conditionally wait for a backing_dev to become uncongested or a pgdat to complete writes
1105  * @sync: SYNC or ASYNC IO
1106  * @timeout: timeout in jiffies
1107  *
1108  * In the event of a congested backing_dev (any backing_dev) this waits
1109  * for up to @timeout jiffies for either a BDI to exit congestion of the
1110  * given @sync queue or a write to complete.
1111  *
1112  * The return value is 0 if the sleep is for the full timeout. Otherwise,
1113  * it is the number of jiffies that were still remaining when the function
1114  * returned. return_value == timeout implies the function did not sleep.
1115  */
1116 long wait_iff_congested(int sync, long timeout)
1117 {
1118         long ret;
1119         unsigned long start = jiffies;
1120         DEFINE_WAIT(wait);
1121         wait_queue_head_t *wqh = &congestion_wqh[sync];
1122
1123         /*
1124          * If there is no congestion, yield if necessary instead
1125          * of sleeping on the congestion queue
1126          */
1127         if (atomic_read(&nr_wb_congested[sync]) == 0) {
1128                 cond_resched();
1129
1130                 /* In case we scheduled, work out time remaining */
1131                 ret = timeout - (jiffies - start);
1132                 if (ret < 0)
1133                         ret = 0;
1134
1135                 goto out;
1136         }
1137
1138         /* Sleep until uncongested or a write happens */
1139         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1140         ret = io_schedule_timeout(timeout);
1141         finish_wait(wqh, &wait);
1142
1143 out:
1144         trace_writeback_wait_iff_congested(jiffies_to_usecs(timeout),
1145                                         jiffies_to_usecs(jiffies - start));
1146
1147         return ret;
1148 }
1149 EXPORT_SYMBOL(wait_iff_congested);